AI
Google's Aachen Loss Is a Reason to Audit Client AI Summaries
Edward Ilin10 min read
An AI answer can be false even when every fact inside it is true. Two German courts have now treated AI search answers as the platform's own statements, not as neutral summaries of other sites. For agency content leads, line-by-line accuracy is no longer a sufficient standard for AI-assisted client work.
True facts can still add up to a false answer.
You approve a summary. Every sentence checks out against its source. The client still ends up described as something it is not. A court in Aachen has just put that exact failure in front of a judge.
Here is what the rulings say and how to protect your client work this week.
The question is no longer only whether an answer is wrong. It is who answers for a misleading answer built from correct parts. Courts are starting to hold the platform responsible. Expect clients to ask the same of you.
Accurate facts can still produce a false AI answer, and courts are now asking who pays when they do
What changed: a German court treated an AI answer as Google's own statement
In a decision dated Sept. 24, 2026, the Regional Court of Aachen ordered Google to stop spreading three versions of an AI Mode answer about a local doctor, according to Search Engine Journal's report on the case. Searchers asked why he left his post as chief physician. AI Mode answered with a serious criminal allegation.
The proceedings it described were real. They involved a different man, also a former university professor. The court could confirm the mix-up from its own records.
The doctor explained the error to Google in late August. Google said in mid-September it would take no action. The court then ordered relief on an emergency basis, with fines of up to 250,000 euros per violation. Google can still challenge the order, as the same report notes.
The reasoning matters more than the outcome. Aachen relied on an earlier Munich ruling that held Google directly liable for AI Overviews linking two publishers to fraud claims that appeared in none of the linked sources. Google argued users could check those sources themselves. The Munich court rejected that, describing the AI's output as "independent, new, and substantive statements" built from third-party content.
Why this lands on your desk: "we only summarized the facts" is no longer a safe defense
The mechanism is entity conflation. A system retrieves documents that match a name plus surrounding attributes such as role, institution, or a recent event. It then writes one fluent narrative from whatever it found. When two entities share enough attributes, the output fuses them.
Your workflow follows the same pattern. An AI-assisted brief, a competitor comparison, or a "who is this executive" bio retrieves, blends, and writes with confidence. If the platform can be held responsible for a blended answer, a client or a third party can reasonably ask the same of the agency that published one.
No court has held that yet. The direction is clear enough to plan around.
- Two companies with similar names in one industry are a conflation risk.
- A franchise location can get confused with its parent brand.
- An executive who shares a name with someone else in the same field is exposed.
The rules are unsettled, so set your own accuracy standard
Do not wait for a clean legal answer. A Berlin court reached the opposite view in a June trademark dispute. It treated AI Overviews as a new kind of search result that bundles information from other sites, not as Google speaking for itself. The Aachen and Munich cases involved invented connections no source made. Berlin's involved accurate repetition of what other sites said. Even so, courts in one country have described the same product in incompatible terms. No higher court has ruled, because the Munich case, headed for appeal, was settled, per Search Engine Journal's account.
The US picture is also in motion. The Washington Legal Foundation has a case page on AI accuracy and federal deceptive-practices law. The Digital Chamber has published a response to the FTC's proposed AI accuracy policy that disputes its premises. Regulators and industry are arguing over the standard right now.
When the referee has not decided the rules, you write your own. Make accuracy a promise you make to clients, not a checkbox someone else defines.
Where overextended content teams are most exposed
Speed is the pressure point. When one lead covers several accounts, the check that gets dropped is the read-for-impression. Each claim gets verified. The whole does not.
The riskiest deliverables combine facts about a named subject:
- AI-assisted summaries of a client's history, leadership, or news coverage.
- Comparison pages that place a client beside similarly named competitors.
- Bios, "about" copy, and press boilerplate generated from scraped sources.
Trust is fragile too. A Raptive survey of 3,083 U. S. Consumers found that when an AI answer conflicts with a source they already trust, only 19% said they would believe the AI answer. Your clients' reputations lean on trusted sources. A misleading AI-assisted piece spends that credibility.
Models are improving. Search Engine Land reported that Google's AI Overviews answered a standard factual benchmark correctly 91% of the time in February, up from 85% in October, per a New York Times analysis with Oumi. A 91% score still leaves a meaningful share of wrong answers. And a benchmark does not measure misleading impressions.
What to do this week
Start with the one client account that is most exposed. Run three steps there, then repeat across the rest.
- Read for impression. Before sign-off, read the finished text once as a stranger would. Write one sentence on who or what the subject is. If it says anything the client would not say about themselves, trace it back to its source.
- Check for lookalikes. For any piece naming a person or company, search the name with their city and industry. Scan the first page for anyone else who fits. Two minutes catches most conflations.
- Tell the client who reviews. Send a short note naming the reviewer, saying you check facts and overall impression, and promising a quick fix if something slips.
If your drafting runs through a platform like Brainpercent, which generates social posts, SEO articles, and videos from one idea, the same logic applies. Faster output makes the impression check more important, not less.
German Courts Now Treat AI Answers as the Platform's Own Speech: Your Questions Answered
Can a client actually sue my agency over an AI-assisted piece that was factually accurate?
Nobody has tested that yet. The German rulings were about a platform, not an agency. What they do show is that a court looks at the overall impression a text creates, not only at whether each sentence matches a source. Carry that into your own work. A client described as something it is not has a complaint whether or not you can point to a footnote for every line.
The more likely path is a retainer conversation, not a lawsuit. A client sees a comparison page that places them next to a similarly named competitor, or a bio that borrows a detail from someone else with the same name, and asks who signed off. A written answer to that question protects the relationship long before any legal theory matters. If a contract or insurance policy covers your content work, check whether it mentions AI-assisted drafting, and ask a lawyer rather than guessing.
What does a "read-for-impression" check look like when I have twelve deliverables due this week?
Keep it to a few minutes and run it on the right pieces only. Skip the generic blog posts. Apply it to anything that names a person or company: bios, leadership summaries, competitor comparisons, press boilerplate. Read the finished text once as a stranger would, then write down in one sentence who or what the subject is. If that sentence contains anything the client would not say about themselves, trace it back to its source.
For the lookalike check, type the subject's name into a search engine along with their city and industry, and scan the first page for anyone else who fits. A shared spreadsheet with one row per flagged deliverable and a column for "lookalikes found" gives you a record without adding a new tool or meeting.
The Berlin and Aachen courts disagree. Which one should I plan around?
Plan around the stricter reading. It costs you almost nothing. If Berlin's view wins, you lost a few minutes per deliverable. If the Aachen and Munich view spreads and you skipped the review, you are the one explaining a mix-up to a client. The cases also involved different problems: invented connections on one side, accurate repetition of other sites on the other. Your own risk sits mostly in the first category, because blended summaries are where invented connections tend to appear.
The US debate is moving too. The Digital Chamber has published a response to the FTC's proposed AI accuracy policy that challenges its premises. The Washington Legal Foundation frames the issue through federal deceptive-practices law. If you serve US and German clients, the rules will stay inconsistent for a while. Set one internal standard and apply it everywhere.
What should I actually tell clients in writing about how we review AI-assisted content?
Write five or six sentences a busy marketing director will read to the end. Name who reviews each piece (a person, not "our team"). Say they check both individual facts and the overall impression. Say that pieces about named people or companies get an extra check against lookalikes. Offer a way to flag a problem and promise a quick correction. Do not promise zero errors. You cannot keep that promise.
A short note in the statement of work or onboarding email beats a new policy document. Clients want to know someone is accountable, and a plain description of your process does that. If a client pushes back on the extra review step, share the Aachen case: the underlying proceedings were real, they simply belonged to a different man, and the court could confirm that from its own records.
Will courts end up writing the real rules for AI accuracy?
Commentators are already saying so. One analysis on Captain Compliance argues that US courts may end up acting as de facto AI regulators, with individual liability cases setting standards before legislation does. The German decisions fit that pattern. They came from individual plaintiffs, not a new statute.
The takeaway for you is timing. Rules built from case law arrive unevenly and shift on appeal. Waiting for clarity means working without a standard for a long time. Build your review step now, refine it as decisions come in, and update your client-facing note whenever a new ruling changes the picture.
What German Courts Now Treat AI Answers as the Platform's Own Speech Means for Your Review Process
The shift is simple. Aachen and Munich both looked at an AI answer and saw the platform speaking for itself, not passing along other people's words. Berlin disagreed, and no higher court has settled it. So stop leaning on "every line checked out" and ask whether the whole piece leaves the right impression of a named person or company.
The payoff is a review habit that fits a crowded week: one extra question at sign-off, a quick lookalike check on anything describing a named subject, and a short written note to clients about who reviews AI-assisted work. Faster drafting, including through a platform like Brainpercent, makes that last check matter more, not less.
Pick your most exposed client account today and run the three steps there first. Once that account is clean, repeat the same pass across the rest of your roster.
Ready to automate all this? Brainpercent is the all-in-one content platform that generates SEO articles, social posts, and videos for you, on autopilot. Start your free trial or see pricing.
Frequently Asked Questions
Make content like this for your brand.
Posts, images and videos from your own site, in minutes.
Try it free